Every technological cycle begins with scarcity. Compute was scarce. Distribution was scarce. Expertise was scarce. Then the cost curve changed—and what looked like a permanent advantage became infrastructure.
Foundation models are moving along the same curve. Capabilities that once required a research lab can now be accessed through an API, adapted in weeks, and reproduced by well-funded competitors. This does not make AI companies less valuable. It changes where value must be built.
Intelligence is becoming infrastructure
As model performance converges, a company cannot rely on raw intelligence alone. Features diffuse. Interfaces are copied. Model access broadens. The strategic question moves from “What can the model do?” to “What can this company uniquely become?”
When a capability becomes abundant, advantage migrates to the system that surrounds it.
The strongest businesses use intelligence as an input—not as the entire product. They turn each customer interaction into better context, each workflow into deeper distribution, and each operating decision into organizational learning.
Four moats that can compound
Proprietary context
Not data in the abstract, but permissioned, structured context that improves decisions and is difficult to recreate outside the customer relationship.
Embedded distribution
A product becomes durable when it sits inside a critical workflow, earns trust, and reaches the next user through the value created for the current one.
A compounding system
The best products improve through use: more feedback sharpens the system, better outcomes deepen adoption, and deeper adoption creates more feedback.
Execution velocity
Speed is not frantic shipping. It is the organizational ability to observe clearly, decide coherently, and translate learning into product faster than others.
What looks defensible—but is not
A thin interface over a broadly available model may be useful, but usefulness alone is not defensibility. Nor are temporary prompt techniques, a long feature list, or a short-lived performance lead. These can create a starting point; they rarely create an enduring destination.
The same is true of data without rights, structure, or a feedback loop. Volume does not equal advantage. A durable data moat must be generated through genuine product value, legally usable, continuously refreshed, and connected to outcomes that matter.
Five questions we ask founders
- 01What becomes stronger every time a customer uses the product?
- 02Which part of the workflow would be genuinely painful to remove?
- 03What do you learn that a model provider or fast follower cannot see?
- 04How does trust accumulate—and how is it protected?
- 05If model capability doubled tomorrow, would your position strengthen or disappear?
The model is the beginning, not the moat.
Intelligence will keep getting cheaper. Building an enduring company will not. The founders who matter most will use this abundance to create systems that know more, serve better, distribute naturally, and improve with time.
